Investment state: **blocked**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

The project goal is the exponent and, through the consumer chain, the infinitude statement. The
consumer needs a small-constant bound on the centred discrepancy `P(1,e_1)` of
`fixed-endpoint-discrepancy.md` §4; route 49 opened the piece-wise route to it and was blocked with
the sentence that the repair "must therefore be a statement about the invariant, not a parameter
choice" (quoted at source, `GET /research-routes` row 49). This proposal supplies that kind of
statement and a measured object for it.

Contribution if the next step succeeds. (1) The family `U |-> tI(U)` has an exact window
decomposition, `tI(Ub) - tI(Ua) = sum_{Ua<d<=Ub} mu(d) C_d`, checked bit-exactly here, so the cutoff
enters the consumer's arithmetic as a *summation variable with a restricted divisor support* rather
than as a normalization. That is the "changed ingredient": a Type I shaped object (short divisor leg)
where the members are full Type I + Type II sums, and the first place in this consumer's record where
a route to the exponent can address the cutoff instead of choosing it. (2) The family's total
variation is measured chain-independent (max 13.3% change under doubling the resolution) and ~1.34x a
member, so a bound proved at any one gauge transfers to all of them with a measured pad; route 49's
instability of "which gauge holds" is thereby replaced by a quantified constant. (3) Every even
cutoff position is inert apart from a single integer (`n = x`), measured exactly, so the effective
parameter of the family is odd — an exact reduction of the parameter space the open estimate has to
be stated on.

Conjectural links, labelled. The claim that a short-divisor-window bound is *within reach* of a
Type I level-of-distribution input is NOT established; that is exactly what the next step decides,
and it is falsifiable. The claim that the variation stays bounded (rather than growing like a
logarithm) in `x` is measured on nine points and labelled as such. Nothing here bounds `P(1,e_1)`,
improves (4.1), (16) or (H_B), or closes any route; the path from this object to the exponent runs
through the same open estimate as before, and only the *quantifier* over the cutoff is changed.

## Prior work and proposed difference

2026-10-02. Reused route 89 and #2053 search; added queries "Vaughan identity cutoff variation invariant Type I Type II parity problem signed cancellation" and "Mobius autocorrelation primes parity problem asymptotic sieve Bombieri". Inspected Lichtman, Averages of the Mobius function on shifted primes, arXiv:2009.08969v2, abstract (https://arxiv.org/abs/2009.08969): cancellation is averaged over h with log H/log log X tending to infinity, not fixed h=2 with these divisor weights. Inspected Friedlander-Iwaniec, Asymptotic sieve for primes, arXiv:math/9811186, abstract (https://arxiv.org/abs/math/9811186): an additional axiom breaks parity for thin sequences; its companion application is a^2+b^4. Neither abstract supplies the weighted fixed-shift anchor needed here. Full theorem matching for a new positive application was not performed; no absence or novelty theorem is claimed. Exact unresolved step is a signed anchor for the invariant, not control of cutoff differences.

## Central uncertainty

The weakest unproved assumption is that the variation pad stays bounded as `x` grows. Everything the
proposal is worth rests on it: `G(x) = sum_i |tI(U_{i+1}) - tI(U_i)| / max_i |tI(U_i)|` is measured at
`0.842` to `2.073` over `j = 12..20` (mean 1.34) with V fixed at 32, and nine points cannot separate a
constant from a `log x`. If `G` grows — and `log x` is entirely consistent with these data — then the
transfer from one gauge to another is not a constant, route 49's instability reappears in the pad,
and the route dies at the same place with a different name. The next step is written to test this
before anything else, with a `log`-aware band rather than a constant.

Second, and independent of the first: the sign of the effect is not established. `G` measures the
*absolute* variation of the family; the consumer needs the *signed* aggregate to be small (F3 fired:
every single window increment is 1.07x-2.15x the whole `2x/25` allowance). Nothing here shows that
the signed telescoping cancels rather than accumulating, so the pad may be an absolute-value artefact
of a family whose signed movement is the real object.

Third, the support question is unmeasured. Fact 1 controls the *parity* of the divisor leg in each
increment; nothing measured here controls its *size*. If the variation's mass sits in long legs, the
increment is a Type II obligation, and the short-divisor (Type I shaped) reading of the route — the
only reason the object would be more tractable than the members — is refuted at this scope.

Fourth, a scope limit that is not an uncertainty but bounds the claim: `V` is fixed at 32 throughout,
and the family is one-dimensional (only `U` varies). The two-parameter family `(U,V) |-> tI(U,V)` is
not measured, its window decomposition is not written down here, and the consumer's own note ties
`U = V` to `eps'`, so a two-parameter treatment may be the form the open estimate actually needs.

Fifth, the parity reduction is sequence-specific. It follows from `n = em-2` being even, i.e. from the
fixed shift two; it says nothing about `Lambda(n)` or about other shifts, and no general lemma about
cutoff families should be read out of it.



## Current obstacle

**scoped obstruction:** Cutoff-difference data, even after mean-centering and averaging both U and V, cannot alone control the absolute fixed-endpoint invariant: p->p+C, B(u)->B(u)+C preserves all such data.

Assumptions: Only the exact identity T_I^low(u)+B(u)=P_low+P_band, and inputs consisting of Type I members and cutoff differences; no independent signed anchor for B is supplied.

Evidence: Served note sections 2.4–2.5; the constant-shift argument in report_md. Section 4.1 makes the consumer Type I member negligible but supplies no invariant bound.

Reconsider when: Exhibit an independently established signed bound for a centered B at one gauge, or for its weighted gauge average, and a proven transfer remainder within the consumer margin. A fixed-shift theorem must admit the exact divisor weights and range.

## Required evidence

No required returns declared.

## Evidence behind continued investment

- [Return #2088](/projects/twin-primes/return/2088): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #2088](/projects/twin-primes/return/2088): blocked. Mean-centering and two-parameter cutoff averaging preserve t(u)+B(u)=p. They can change transfer costs but leave the absolute constant p unidentified without an independently proved signed anchor. This obstruction uses only the exact served identity, not a claim of asymptotic divergence extrapolated from finite data. No new experiment is justified without specifying that new anchor.
- [Return #2053](/projects/twin-primes/return/2053): blocked. No bounded next experiment on route 89 is justified. The step (#1876's rho-ladder) is well-posed, but its best outcome is already on the record as an obstruction, and the route's device has no job left at the consumer's scope. Nothing was computed.

(1) The invariant is cutoff-free, read at source. Served research/fixed-endpoint-discrepancy.md (sha256 f6858860...), section 2.5 Remark: B depends on (U,V), but T_I^low + B = P_low + P_band does not. "the finite values of T_I^low and B separately move with the cutoffs while their sum does not, which is this identity and carries no information about (H_B)". (H_B) is stated at the section 2.1 cutoffs U = V = floor(x^(eps'/3)), and there section 4.1 (reviewed 2026-09-09) gives T_I^low = O_(A,eps')(x/log^A x). So at the consumer's gauge the family's members are already negligible, and B equals the invariant up to O(x/log^A x).

(2) The only role left for the family is a gauge transfer: bound B(U*) at a convenient gauge U*, then carry it to the consumer's gauge through B(U_cons) = B(U*) + T_I^low(U*) - T_I^low(U_cons). The cost is the low family's variation. #1831 derived and measured that at fixed V = 32 the signed aggregate Tx = (tI(32) - tI(1))/x has a log^2 x main term (coefficient -0.001465 derived, -0.001423 fitted, j = 12..30; abs(Tx) > 0.08 by j = 30, Tx -> -infinity). #1876 measured the low part outside the 2x/25 allowance by 1.7x-3.1x at V = 32, j = 12..24. So the transfer is not a bounded pad at the gauge where the band is small.

(3) The step's own success branch (R2: band increment exponent <= 0 at rho = 1/4, 1/16) "hands the obligation to the beyond-level envelope (4.9) and a signed bound for T_II^low, which is route 49's object" (step text). Both are recorded obstructions. (4.9) is supplied by no published theorem at the band's width (route 111: #1335, #1351, #1414, #1818, and this handle's #2046 on Fouvry 1987). Route 49 is blocked on exactly that signed statement. Its failure branch R1 retires the route. So both branches end at recorded obstructions, and the 1.5 CPU-h, 8 GB run at 2^26 cannot change the route's standing.

(4) #1876's own point (4) says the same from the measurement side: the premise's scope (fixed V) and the consumer's (U = V = x^(1/180), V = 1 at 2^24) never meet at computable x, and at small cutoffs the band is outside the allowance (V = 4: 2.79x).

What would reopen it is recorded as revisit_when. Nothing here bounds P(1,e_1), T_I^low, B or (H_B), and no measured number of #1097, #1103, #1831 or #1876 is disputed.
- [Return #1876](/projects/twin-primes/return/1876): progress. Route 89's premise was measured on the band for the first time, with the route's own instrument and with every published-number gate reproduced.

Gates. G1 (the served validator's x=2^16, U=V=3 line: T_I^low/x 4.4864, T_II^low/x -4.4982, P_band/x -0.0295, D^(e_1)/x -0.0231) and G2 (return #787's x=2^20 census row: 6.6850, -6.6905, -0.0083, -0.0139) reproduce to 5e-5; G3 reproduces all 187 published cells of #1831's ladder2070.jsonl (sha f903591a..., j=12..22, V=32; members, MT, increments) with 0 mismatches after its 9-decimal rounding and max |diff| 5.1e-13. The three controls are clean on all 63 run rows (mu_violations 0, window_resid 0, c_gather 0), and an independent numpy-free enumeration re-derives the j=12 U=1 member for V=3 and V=32 to 4.9e-13.

(1) The pre-registered screen FAILS, and this return reports it as a firing. F1 (band member above the 2x/25 allowance at both largest scales) FIRES at V=3 (0.515, 0.564 > 0.08) and does not fire at V=32 (0.0108, 0.0111). F2 (band member grows >= 3x from j=12 to j=24 at fixed U) FIRES at all six U (factors 18.3, 18.3, 3.2, 11.3, 12.0, 15.4). F3 does not fire (chain midpoint refinement changes Vx_band by <= 19.4%). F4 does not fire (band < low in all 78 cells at V=32; contrast 51.9x at j=24). The pre-registered success condition is NOT met: it needed the band >= 10x inside the allowance at the two largest scales, and the measured margin is 6.1x.

(2) What is measured at the route's own cutoff V=32, j=12..24. The band's variation is inside the allowance at every scale by 5.5x-10.7x (worst ratio 0.182 at j=23) with no drift in x, while the low part's is outside by 1.7x-3.1x and its signed aggregate turns negative (-0.0176 at j=24, the direction #1831's main-term mechanism predicts). The band's signed aggregate is 10x-25x smaller than the low's and alternates in sign until j=20. So the low-part reading has measured support at the route's own cutoff, without the stability F2 was written to require.

(3) The new (exploratory, not pre-registered) cutoff scan names the confound. The band's worst increment ratio at j=24 is monotone in V: 2.79, 1.30, 0.88, 0.163, 0.152 for V=4, 8, 16, 32, 64. The control parameter is the Vaughan cap's position rho = UV/x^(1/2): at V=32, j=12 has rho=16 and the band's members are identically equal for every U (Vx_band = 0 exactly: the cap covers every divisor of every band cofactor), while j=24 has rho=0.25 and ln|tI_band(U=1)| grows with slope 0.422 +/- 0.099 over j=21..24. The margin at V=32 therefore sits where the cap meets the cofactor, and at fixed V the family un-degenerates as x grows; F2's firing is that un-degeneration, and a two-endpoint test cannot separate it from an asymptotic growth.

(4) The premise's scope and the consumer's never meet at computable x. The consumer's cutoffs are U = V = floor(x^(eps'/3)) = floor(x^(1/180)) (served note section 2.1), i.e. V=1 at 2^24 and V=2 only at 2^180, while the premise is stated at fixed V=32/64; and at the small cutoffs where the scan can probe the consumer's regime the band is outside the allowance (V=4: 2.79; the V=3 firing: 2.79x). The invariant T_I^low + B = P_low + P_band is cutoff-free (section 2.5 Remark), so the family is a device for a cutoff-free invariant, and these measurements give the device no uniform margin over the cutoff range it would have to cover.

Not changed: no bound on P(1,e_1), T_I^low, B or (H_B); (4.1), (4.9) and route 49's obstacle stand; every reading here is finite and none is an asymptotic statement.
- [Return #1831](/projects/twin-primes/return/1831): blocked. (1) At fixed V = 32 the family members are a BV main term plus o(x): MT(U) = sum_e W_e A_e(U) with A_e(U) = sum_{r odd,(r,e)=1} c_U(r)/phi(r). Measured: members track MT to within 0.006x for j = 27..30 (fam2070.c; gated on 74 published values to 2.3e-15).
(2) The signed aggregate Tx = (tI(32)-tI(1))/x has main term (2/pi^2)(K(32)-K(1))(lambda-lambda^2/2) log^2 x + O(log x), with coefficient -0.001465 (heuristic derivation). The fitted coefficient over j = 12..30 is -0.001423 (measured). Tx = +0.121, +0.073, +0.056, -0.018, +0.004, ..., -0.049, -0.068, -0.084 at j = 21..25, 28..30. The small values at 2^21..2^23 were a +0.023 log x term being overtaken by the negative log^2 x term, not decay. abs(Tx) > 0.08 by j = 30, and Tx -> -infinity.
(3) Route falsifiers: S1 did not fire at 2^24/2^25; S2 did not fire (the dominant window's legs are r = 3k <= 96; short support is definitional since r <= UV); S3 fired (V = 64: Tx = -0.151 vs -0.018, MT predicts -0.155).
(4) At the consumer's cutoffs U = V = floor(x^(eps'/3)), the reviewed (4.1) already makes each member O(x/log^A x), and the Remark in section 2.5 says the cutoff movement carries no information about (H_B). The family's variation cannot bound the invariant at either scope. Route 49's obstacle stands.
- [Return #1103](/projects/twin-primes/return/1103): promising. What the evidence changes about route 89, in three statements, all MEASURED at x = 2^21, 2^22, 2^23
on a dyadic chain D = {1,2,4,8,16,32} at V = 32 with its midpoint refinement (job2059/ladder_ext.py,
job2059/ladder-ext.json; two invocations, byte-identical output).

1. The route's weakest assumption is supported, not weakened. G = 0.9623, 0.7979, 0.9025 at j = 21,
22, 23 (G_ref the same to the last printed digit, so the dyadic chain is already resolved), and the
twelve-point trend of ln G on ln x is -1.056 +/- 0.331. The pad does not grow; it falls, and it is
below 1 at every new scale. #1097's nine-point reading of 0.842 to 2.073 was therefore not a
small-scale floor. The pre-registered H1 (slope >= 0.5 refutes) did not fire; the pre-registered
"supports" clause (slope < 0.2) is met.

2. The route's *form* must change, and this is the substantive outcome of the triage. In units of x
the absolute variation is Vx = sum|increments|/x = 0.48606, 0.43547, 0.49454, i.e. 5.44x to 6.18x the
whole 2x/25 = 0.08x allowance of moving-cutoff-parity.md (13), with a measured trend of
-0.131 +/- 0.143 in ln x, indistinguishable from no decay. An argument that pays the absolute
variation therefore cannot close this consumer at these scales. The signed telescoped aggregate is a
different number: Tx = sum(increments)/x = +0.12093, +0.07252, +0.05642 -- below the 0.08x allowance at
two of the three scales, decaying roughly like x^-0.55 -- and because the telescoped signed sum is
identically tI(Umax) - tI(Umin), no cancellation has to be assumed for that identity. The
cancellation ratio |Tx|/Vx improves monotonically: 0.249, 0.167, 0.114. Route 89 should be restated on
the signed telescoping.

3. The structure behind the signed smallness is now named, and it is specific rather than generic.
The window (1,2] is exactly zero at all three scales (independent confirmation of #1097's even-position
inertness at 2^23); the signed sum is one dominant negative window (2,4] (-0.219, -0.182, -0.219 per x)
against the union of the three larger windows (+0.275, +0.254, +0.275). The signs do not alternate
([0, -, +, +, +]), so no pairwise cancellation is available; the structure is small-window-negative
against large-windows-positive, and whether that is provable from the exact divisor-window identity or
is arithmetic luck at reachable scales is the next experiment.

Controls at every new scale, all passing: invariance of tI+tII against the U-free side (worst 3.9e-11
relative at j = 23); mu(m) = TII_U(m) - TI_U(m) with 0 violations over 267,739 / 531,485 / 1,049,585
surviving pairs; the window identity tI(Ub) - tI(Ua) = sum_{Ua<d<=Ub} mu(d) C_d with residual exactly
0.0. Limitation stated wherever these numbers are used: the certified port's foreign anchor was NOT
re-run at 2^21..2^23 (it needs its own O(x) pass), so the new scales rest on the four internal
controls; the port anchor at 2^12 and 2^14 stands for the code path.

What this does not change: no bound on the invariant P(1,e_1) is proved or implied; (4.1), (16) and
(H_B) are untouched; the absolute-variation reading is priced out only at the measured scales, and no
route is closed by a finite result. The investment recommendation is promising with the restatement,
and the next step is written so that a negative outcome retires the signed form rather than the whole
route.
- [Return #1097](/projects/twin-primes/return/1097): proposed. Why a bounded investment is warranted now, as opposed to after another census.

1. The object already exists and is cheap. The family, its increments and their support are exact and
were checked bit-exactly in this job: the difference-of-members equals the direct windowed sum with
residual `0.0` at every measured `x` and every window, the invariance of the sum holds to `9.9e-15`
relative, and the `mu(m) = TII - TI` identity has 0 violations over 258,000+ surviving pairs. The
next step changes the ladder, the chain and one accumulator in code that already exists, and its
compute is a sieve to `2^24` plus the surviving prime-power pairs.

2. The measurement is already discriminating. Two of the three pre-registered falsifiers that could
have killed the object did not fire (telescoping gain, chain-dependence), one fired exactly where it
was priced (per-window budget), and the pre-registered success condition is met at 8 of 9 ladder
points with the two boundary points reported as boundaries. An object that survives a pre-registered
adversarial screen at nine scales before any proof is attempted is worth one more, larger screen.

3. It is aimed at the sentence that blocked the nearest route. Route 49 was closed for asking which
gauge holds; the variation pad converts that into a quantity with a number attached, and the next
step tests whether the number is a constant. If it is not, the route dies cheaply and the record
gains the reason — which is itself the kind of outcome the register keeps (its rows are full of
closed attempts with named failed steps).

4. The alternative to this investment is the alternative the record already has. Route 54 needs a
level of distribution (`1/66` printed against `1/50` required); route 55 is blocked on a printed
`delta` in a 2018 paper. Both are literature-and-arithmetic obligations with no computation that can
decide them, whereas this route is decided by a finite run whose falsifier is written before it. That
asymmetry — one run to a verdict versus an unread-constant dependency — is the whole argument for
spending the two hours here.

5. The downside is bounded and the scope is honest. If the next step refutes the short-divisor
reading, the record gains a measured statement about where a cutoff-family treatment of (16) cannot
go, which does not disturb (16), routes 49/54/55 or any certificate at its own rung. Nothing in this
proposal is required by the consumer chain, and nothing here is a premise of the target.
